The AMD CTO's "next three years" for optics now has a date: the 2027 MI500 gets an optical option, and copper and optics each hit a yield wall

Key takeaways

In January, AMD's chief technology officer, Mark Papermaster, still answered the question of when optics replaces copper with a rolling window. "Think of photonics as coming on board in the next three years, all the pieces of the supply chain coming together, starting at just the biggest clusters," Papermaster told More Than Moore. The trouble with a sentence like that is that it never comes due. Said in January, it means three years out. Said again in July, it still means three years out, and no supplier can stock parts against it.

AMD's July 23 event pinned that window down. AMD's official press release states in writing that the next-generation MI500 series of GPUs launches in 2027. On stage, AMD confirmed that the MI500 generation supports both copper and optical links, while the Helios rack launched that day stays all-copper. This is the first time a GPU maker has put optics on the roadmap of a named product generation. NVIDIA's matching marker is 2028. "When do optics arrive" is no longer a question that always sits three years out. The answer is now a marker that can be checked against the 2027 launch, and it can turn out wrong.

Three findings go further than the first version of this call. First, "optics is delayed" and AMD's marker turn out to describe different populations. The delay is about co-packaged optics in switches sold on the open market: one switch carries 32 optical engines, and volume waits on an industry standard and a second supplier. AMD's marker is about an optical option inside AMD's own GPU generation: one vendor, vertically integrated, form not yet decided, possibly limited to the largest configurations. These are two different populations, and one sentence about "optics delays" doesn't cover both. Second, the claim "copper wins near-term" was tested against the marker, and it holds only halfway. Copper does win this generation. But volume production of NVIDIA's next copper rack has slipped to 2028, because a 78-layer circuit board in that rack can't be made at acceptable yield. In the densest layer, GPU to GPU inside a rack, "optics delayed, copper advances" turns out to be optics late and copper late too. Third, the conditions that would prove this wrong are written down with dates. The nearest marker falls in the fourth quarter of 2026, the deadline an industry group set itself for a first optical specification. The heaviest falls in 2027: whether optics actually shows up when MI500 launches, and in what form.

One event turned a rolling window into a fixed date

What happened on July 23 is worth splitting into four layers, because each layer of the commitment has a different strength and supports a different kind of judgment.

Layer one: the year, committed in writing. AMD's official press release says: "Next-generation AMD Instinct MI500 Series GPUs are coming in 2027." The same release adds: "AMD Instinct MI600 Series GPUs are coming in 2028." It also defines the next rack in full: "The next-generation AMD Helios 500 rackscale solution will be powered by AMD Instinct MI500 Series GPUs and AMD EPYC 'Verano' CPUs, with next-gen Pensando 'Como' and 'Monza' networking." Pensando is AMD's networking chip unit. So the year is a commitment made at press-release level.

Layer two: optics, on a spoken roadmap. The word "optical" appears nowhere in the press release. The written text stops at next-generation interconnect technology. Optics shows up on the keynote roadmap. The trade outlet SDxCentral reported it precisely: "AMD's MI500 series graphic processing units (GPUs) will cater to both copper and optical connectivity, the chip giant confirmed." StorageReview's account of the roadmap slide matches: "the CDNA 6-based MI500 series arrives in 2027 with next-generation HBM plus copper and optical interconnect." The key word is "and." It is copper and optics, not copper replaced by optics. Optics is an option in this generation, not what defines it.

Layer three: the form, which AMD declined to give. What form will the optics take? Co-packaged optics, with the optical engine packaged right next to the chip; optical links inside the rack; or pluggable optical modules? Krishna Doddapaneni, a senior vice president at Pensando, was asked. Doddapaneni called it a "future thing," said AMD has been discussing it internally, and declined to share more, as SDxCentral reported. The distinction is not a technical detail. Co-packaged optics benefits the packaging and optical-engine supply chain. In-rack optical links benefit cable and transceiver makers. The two lists of beneficiaries barely overlap. In April a supply-chain rumor said GlobalFoundries would make the photonic chips for MI500's co-packaged optics and that ASE, a major chip packaging and test house, would package them. That rumor came three months before the event and has had no official confirmation since. Its evidence grade has to be kept apart from the two layers above.

Layer four: suppliers, blank. AMD has not named any product-level partner for MI500's optics. What can be found are peripheral signals. AMD Ventures and NVIDIA are both strategic shareholders in Ayar Labs, a startup building co-packaged optical links; Ayar closed a US$500 million Series E in March and said the money goes to scaling volume production and test capacity. In January Papermaster named Celestial AI, which uses optics to connect chips to memory. In June, Oriole Networks, a startup building photonic networks for GPU clusters, announced a collaboration pairing its technology with AMD GPUs. That makes three threads, and none of them answers who makes MI500's optics.

Laying out the four layers also records a correction to this call. When the call first appeared in the August 16 daily brief, it had two structural weaknesses. All the event detail came from one analyst, Ian Cutress, relaying the event in posts on X. And "the window collapsed" rested on stitching the CTO's January remark onto July's roadmap. Both weaknesses are now resolved. The official press release, plus at least three independent outlets that were in the room (SDxCentral, StorageReview and ServeTheHome), confirm what was shown. "Optics scheduled into the MI500 generation" no longer needs any stitching: the roadmap says so directly. So the year is confirmed, and the single-source weakness is gone. The form and the suppliers remain open, and AMD has given no answer on either.

One more coordinate gives the marker weight: NVIDIA set one too, a year later. NVIDIA's official technical blog describes the interconnect inside the next Vera Rubin rack as still copper. Optics between GPUs is written into the Kyber platform after that, which "will scale up into a massive all-to-all NVL1152 supercomputer using similar direct optical interconnects for rack-to-rack scale-up." SDxCentral's roundup adds that NVIDIA won't offer optics and copper together for both scale-up and scale-out until the Feynman generation in 2028, so MI500's 2027 launch puts AMD "a year ahead." When several players compete, one player's commitment is a signal aimed at rivals and customers at the same time. With both companies on a one-new-rack-a-year cadence, writing optics into a named generation is a competitive weapon, so the commitment deserves a discount. But because it is a weapon, any slip will be seen by the whole market. Competition is what ended the era of the rolling window.

The copper route: it won this generation, but each vendor hurts differently

The incumbent on this branch of the tree is copper. What is being claimed. The densest connections between GPUs, the part the industry calls scale-up, which ties 72 or more than a hundred GPUs inside a rack into one system, run entirely on copper today. Who's betting on the claim. AMD's current Helios generation: StorageReview's hands-on description says "every GPU-to-switch link is copper run through four blind-mate cable cartridges at the rear." NVIDIA's in-rack interconnect through Rubin Ultra. And Broadcom CEO Hock Tan, who said last December, as SDxCentral reported from Broadcom's earnings call, that engineers are still finding ways to "do scale up within a rack in copper as long as possible." The two vendors' evidence is not the same grade. AMD's side is backed by a named customer order. On NVIDIA's side, the evidence in this piece is a platform roadmap, with no matching named customer. How hard the evidence is. Copper is the only thing shipping at scale. Helios is in full production and ships at the end of the third quarter, and Anthropic's commitment to deploy up to 2 GW of capacity is anchored on this all-copper generation.

The ledger behind "copper wins this generation" hides two separate pains, and they land on the two rivals in completely different shapes.

The first is AMD's. A late-July supply-chain field study by the research firm SemiAnalysis found that the Helios ramp is slow. One root cause is that AMD's 200G SerDes, the interface circuits that send and receive a GPU's signals, are weaker than NVIDIA's. In Meta's version of the deployment, about 85% of in-rack links need an extra chip called a retimer to make up for signal loss. That means more than 550 Broadcom retimers per rack, plus 1,728 flyover cables. Each one adds cost, power and a possible point of failure. Copper's physical limit is a floor the whole industry shares: signal loss gets worse with frequency and distance. But who hits it first, and how hard, depends on each vendor's engineering. In the same copper generation, AMD pays a noticeably higher "retimer tax" than NVIDIA. That gives "AMD jumps to optics a year before NVIDIA" a structural motive. The motive is not a technology lead; copper simply costs AMD more. The same evidence cuts the other way too. A company that already finds copper harder than its rival does is about to take on optics, which is harder still. The question mark over AMD's execution has to stay up alongside the motive.

The second pain is NVIDIA's. NVIDIA's next-generation Kyber rack for Rubin Ultra replaces cables with a 78-layer copper circuit-board midplane. SemiAnalysis's assessment, as relayed by Tom's Hardware, is that the board "cannot be manufactured at reliable yields," and the whole Kyber rack slips to 2028 as a result. So while optics challenges copper, copper's own next step has run into manufacturing yield. The market consensus of the first half of this year, "optics delayed, copper advances," looks different in the densest layer: optics is late, and copper is late too.

What would prove this section's reading wrong: if AMD fixes its signal quality in the MI500 generation (or brings out its own switch), the argument that the retimer tax pushes AMD toward optics weakens. If Kyber ships on its 2028 schedule, copper's yield pain was a one-off rather than a structural signal.

The optical route: the yield wall is real, and so are the orders behind it

The challenger is optics. What is being claimed. Packaging optical engines right next to a switch or accelerator (co-packaged optics, CPO), or at least moving the links between racks to light, buys an order-of-magnitude improvement in bandwidth per watt. Who's betting on the claim. NVIDIA has started shipping optical switches to select partners. Broadcom is on its third generation. AMD has just written optics into 2027. An industry group set a deadline for a specification five days before this piece. NVIDIA's shipments are still limited volumes to selected partners, not volume deployments at a hyperscale cloud; the fourth condition under "what would prove this wrong" waits for exactly that. How hard the evidence is is the point of this section, because the evidence on both sides is real.

Start with the wall. SemiAnalysis's June 9 report, "Powered Down, Lights Off: 800VDC Pushout and CPO Delays," is for paying subscribers, and the full text is not public on SemiAnalysis's site. The three figures used here are the ones that three public summaries of the report agree on. The report's yield math has been widely repeated. Assume each optical engine is attached with 95% yield. A switch that carries 32 optical engines then has a full-pass yield of 0.95 to the 32nd power, about 19.4%. Insertion loss on the board may exceed 3.5 dB, close to the budget limit. The report's conclusion: volume production of co-packaged optics may slip to 2028 or even 2029. That month the report set off a sharp sell-off in US optical-communications stocks. That yield is the bottleneck has an independent second source. Morgan Stanley estimates TSMC-side packaging yields at 50% to 60% and downstream assembly yields at 20% to 50%. TrendForce's late-July assessment reads: "High manufacturing yields and effective thermal management are essential, limiting mass production to only a small number of suppliers." TrendForce puts the ramp in 2027 to 2028. ASE was blunter: its chief operating officer told an industry conference in mid-August that the industry has not yet built the shared testing standards and simulation groundwork the technology needs. One figure needs a note on its standing. Earlier circulating numbers such as "TSMC's COUPE yield is only 60% to 65%" exist only in relays of the paywalled report, and no public source carries them word for word. This piece does not treat them as constants.

The supply side points the other way. Line up the supply-side moves of the second half of 2026, and the shape doesn't fit "a big delay." TSMC's COUPE, its silicon-photonics packaging platform, has entered mass production. NVIDIA's Spectrum-X optical switch has begun shipping to select partners, which TrendForce describes as "marking the official transition of co-packaged optics (CPO) into the mass production phase." The most forceful published rebuttal of SemiAnalysis gives an order figure. NVIDIA's guidance to the laser makers Coherent and Lumentum for high-power lasers rose from about 40 million units in January to about 100 million in April and May, a 2.5-fold increase. Lasers are the key part of the external light source in co-packaged optics. Then look at when parts are scheduled. Lumentum has won incremental high-power laser orders worth several hundred million dollars, with shipments in the first half of 2027. The first shipments of co-packaged optics for scale-up are scheduled for the second half of 2027. Marvell's acquisition of the optical-interconnect company Celestial AI closed in February; Marvell's official guidance puts Celestial's first revenue contribution in the fiscal-year period beginning in the second half of calendar 2027. Corning, the glass and optical-fiber maker, won a multi-year partnership and a direct investment from NVIDIA and is building three new plants in North Carolina and Texas. Last comes the standard. On August 13, twenty companies formed a silicon-photonics working group at OCP, the Open Compute Project, where the big cloud buyers and hardware makers write shared open specs for data-center hardware. The group published an architecture white paper of about 300 pages and set the target for its first formal specification at the fourth quarter of 2026. An OCP specification and a finalized industry-wide standard are two different things, though; whether the first counts as the second depends on adoption.

Put the demand-side marker and the supply-side orders in the same table, which was impossible before the July marker existed. The result: laser shipments in the first half of 2027, first scale-up optics shipments in the second half of 2027, Celestial's first revenue in the second half of 2027. All of it lines up with MI500's 2027. The supply chain's orders point to 2027. Meanwhile Broadcom, the largest switch maker in the ecosystem, still talks down optics. Hock Tan says copper should last as long as possible and that on silicon photonics "we're not quite there yet." Yet Broadcom's own third-generation co-packaged optics switch is already shipping to early-access customers. When what a company says and what it does diverge like this, orders are the better guide. None of these supply signals says "for AMD." Most of them revolve around NVIDIA's supply chain. Matching them to 2027 is an alignment of timing, not of customer lists.

Resolving the collision: what is delayed and what is pinned are different things

Now take the contradiction between the two claims head-on. One side says co-packaged optics won't mature until 2028 or 2029. The other is AMD writing optics into its 2027 generation. On the surface that is a two-year gap, which suggests one side must be wrong.

This piece's call: until the populations behind the two sentences are laid out, the contradiction is false. The delay SemiAnalysis describes is about large-scale commercial deployment of co-packaged optics in open-market switches. Yields multiply across 32 optical engines per switch. Hyperscale clouds won't bet big until there is a second supplier, so volume waits for an industry standard. AMD's marker is about an optical option within AMD's own accelerator generation: copper and optics both supported, form chosen by AMD, suppliers tied to AMD, possibly introduced only on the largest configurations. None of the open market's conditions has to be met first. Papermaster's January framing in fact foreshadowed this shape: optics "starting at just the biggest clusters," and products designed for copper and optics together, because "You can architect in flexibility." One is the maturity timeline of an open ecosystem. The other is the introduction date of a single vendor's vertically integrated product. The price of vertical integration is small volume. The vendor trades scale for speed. The second can naturally run ahead of the first. In the car industry, too, new technology usually shows up first in one maker's flagship, long before the parts market standardizes.

This split reading directly rewrites the three possibilities listed two days earlier. The call first appeared, not yet settled, in the August 16 daily brief. Its three possibilities were that AMD's optics might not be co-packaged at all, that AMD might slip, or that SemiAnalysis was too pessimistic. Now they can be ruled on. The third possibility, that SemiAnalysis was too pessimistic, doesn't need to hold. As long as "delay" is limited to large-scale commercial deployment on the open market, and "2027" to an option-level introduction by one vendor, both sides can be right at once. The first possibility, that AMD's optics might not be co-packaged, is no longer just a guess. AMD officially declined to say what form the optics will take. So until the form is disclosed, "MI500 has optics" still can't support any specific supply-chain stock. That discipline was written down in the August 16 issue and holds unchanged today. The second possibility, that AMD might slip, stays open, and it is the marker's greatest value: whether AMD slips will be known in 2027.

The dividing variable is the optical form together with the supply structure. If MI500's optics turns out to be in-rack optical links or a pluggable solution tied to one vendor, the split reading holds, and the fight over co-packaged optics yield timelines has nothing to do with MI500. If it turns out to be true co-packaged optics on an open supply chain, it runs straight into the yield and standards timeline. The two-year gap between 2027 and 2028–2029 then becomes a real contradiction, and one side has to give. That is why OCP's fourth-quarter-2026 specification deadline matters: it answers whether the foundation of the open route is ready a year before MI500.

What this changes, and for whom

• Optical-interconnect procurement and capacity planning: the demand side now has two named markers, 2027 (AMD, option level) and 2028 (NVIDIA, starting rack to rack). Stocking decisions move from loose preparation for a window that is always three years out to staged preparation for two named generations. This piece has shown the method: set visible order and shipment dates (lasers, optical engines, glass, packaging) next to the two markers. Orders placed well before 2027 risk arriving ahead of demand; orders placed well after it risk a shortage. The August reading is "aligned, no contradiction."
• Valuation stories for optical-interconnect companies: "CPO demand" splits into two lines. The first is co-packaged optics in switches, the network that connects a GPU cluster to the outside. It is already shipping, its bottleneck is yield and standards, and its beneficiaries are switch chips, optical engines and external lasers. The second is optics in GPU generations, from 2027, with form and suppliers undecided. Until the form is disclosed, the second line cannot carry a story about any named stock. A pitch for "MI500 plays" rests on knowing which optical form AMD will use, and AMD has not said.
• Buyers of AI racks: "copper and optical both supported" foreshadows SKU differentiation. Within one generation, the largest cluster configurations get optics first and mainstream configurations stay on copper; AMD's CTO laid out that script in January. For multi-year purchasing plans, the open terms are which configuration tier first offers an optical option, at what premium, and with what yield terms in the contract.
• The AMD-versus-NVIDIA contest: once both companies are on a one-rack-a-year cadence, roadmap dates themselves become weapons, and a weapon can feint. So these two markers are best kept as a record to check rather than taken on trust: AMD 2027, NVIDIA 2028, checked against every earnings call and every launch event from now on. The flip side of an annual cadence is a public check every year. In the rolling-window era there was nothing to check against.

Our call

In AI racks, the demand side of the copper-to-optics handover moved in July 2026 from a rolling window to named generations: AMD officially confirmed that the MI500 generation (2027) supports both copper and optics, and NVIDIA put rack-to-rack optics in Kyber, set for 2028. But the commitment comes in four layers: the year is in writing, optics is on a spoken roadmap, AMD declined to give the form, and the supplier list is blank. The supply-side "optics yield delay" does not contradict this marker, because the two describe different populations: the delay concerns large-scale commercial deployment of co-packaged optics in open-market switches, while the marker is an option-level introduction by one vertically integrated vendor. The dividing line is what form the optics takes and what supply structure comes with it, which the MI500 launch in 2027 will reveal. At the same time, "copper wins near-term" has to narrow to "copper wins this generation": AMD pays a higher retimer tax on copper than NVIDIA does, and NVIDIA's next copper midplane has slipped to 2028 on yield. The densest layer of interconnect means GPU to GPU inside a rack, which is a different thing from the rack-to-rack links above. In that layer, copper is fighting the yield of NVIDIA's 78-layer board, and optics is fighting the yield of assembling many optical engines. Laser, optical-engine and Celestial AI shipment dates already line up on 2027. Most of those orders sit in NVIDIA's supply chain, and none is marked as being for AMD, so the match is in timing, not in customer lists.

What would prove this wrong: First, optics is absent when MI500 and Helios 500 formally launch, or is pushed to the generation after. "AMD optics in 2027" is then falsified, and the marker pays off as the first recordable slip, which is part of what the marker is for. The window is calendar 2027, judged by AMD's official launch. Second, the optical form is disclosed as co-packaged optics on an open supply chain. The split reading, "the delay and the marker don't contradict each other," is then void, 2027 versus 2028–2029 becomes a real contradiction, and this piece has to be re-ruled. If the same disclosure shows in-rack optical links or pluggables, the split reading is confirmed, and the market's picture of "MI500 equals a co-packaged optics play" has to be cut back. Third, OCP's first silicon-photonics specification misses its fourth-quarter-2026 deadline, a date OCP chose. The open route's standards timeline then moves back to SemiAnalysis's pessimistic side. Fourth, any hyperscale cloud deploys co-packaged optics in volume, not as a pilot. The "copper wins near-term" reading on this branch of the tree is then void. In the other direction, if yield readings (TSMC's COUPE, ASE's testing) show no public evidence of improvement by mid-2027, "option-level introduction" should be downgraded to "demonstration level." Fifth, either Lumentum's first-half-2027 shipments or the first Celestial revenue at Marvell in the second half of 2027 (both dates set by Lumentum and Marvell) slips. The reading "supply orders are aligned on 2027" then fails, and the mismatch signal flips to the shortage side. The first two conditions take AMD's official disclosures as their source. The third to fifth take each company's self-set schedule as their source. Each will be reviewed every quarter and written into this piece's follow-up record.

This piece updates the ruling on this branch of the tree. The old "copper wins near-term, optics delayed" narrows to "copper wins this generation (Helios and Rubin); the demand side now has two markers (2027 and 2028); and the copper and optical routes each face a manufacturing-yield wall." The tree's original conditions for what would prove it wrong are unchanged, and four new dated markers are added.

Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.